LOVART / Xingliu Market Research

1. Introduction

LOVART (Chinese product name "Xingliu") is a design-focused AI Agent launched by LiblibAI (Beijing Qidian Xingyu Technology Co., Ltd.), officially positioned as "the world's first design Agent". Unlike "generation engines" such as Midjourney and FLUX, Lovart does not train image base models itself; instead, it Agent-izes the entire design production chain of "understanding a design brief → decomposing tasks → routing to suitable image/video models → batch generation → layout and delivery on a unified canvas".

Its product form can be summarized as: the user gives a one-line requirement (Brief), and the Agent delivers a systematic set of design assets (Logo, poster, social media image set, packaging mockups, storyboard, short video), rather than a single image. Among the 16 platforms surveyed, Lovart / Xingliu represents the "design middleware / Agent-ized design" route, and the most prominent layers of its six-layer Harness are L3 (Orchestration and Control) and L4 (Brand Kit as a first-class citizen).

An important note: Xingliu (xingliu.art) has announced it will cease service at 2026-10-10 24:00, and users' works must be migrated to Lovart.art. This fact directly affects the selection decisions of domestic users; see 1.2 for details.

1.1 Basic Information

1.2 Development History

1.3 Positioning in the AI Harness Framework

Lovart / Xingliu's distinctive value among the 16 platforms in this group is that it makes "orchestration" rather than "generation" a first-class citizen of the product:

• The core object of traditional image platforms (Midjourney, FLUX) is "a single generation";
• ComfyUI's core object is "a workflow graph";
• Lovart's core object is "the mapping process from a design brief (Brief) to a complete set of deliverables (Deliverables)".

Corresponding to the six-layer model, its strengths and weaknesses are very clear:

• L3 (Orchestration and Control) is the strongest: task decomposition, multi-model routing, parallel batch generation (about 40 assets per run), and canvas-level layout together form a complete design-production orchestration.
• L4 (Memory and State) has a clear first-class abstraction: Brand Kit fixes Logo, color palette, and font system as asset objects reusable across tasks, with quantities limited by subscription tier (5 / 10 / 30 / 100) — it is the only platform in this research that turns brand assets into an explicit quota resource.
• L1 (Context Engineering) is strong: brief + Brand Kit + multi-round conversational editing form a closed context-iteration loop.
• L5 (Evaluation and Observation) and L6 (Governance and Security) are relatively weak: no public Eval Set; no official documentation on labeling and portrait-rights compliance implementation.

2. Glossary

3. Feature Description

3.1 Design Agent Task Execution Paradigm

Lovart's execution paradigm can be broken down into five phases:

The difference from the traditional "prompt → image" flow is that users no longer need to know which model to use, what prompt to write, or what parameters to set — these are taken over by the Agent. The cost is reduced visibility and controllability of the generation process for users.

3.2 ChatCanvas and the Infinite Canvas

ChatCanvas is Lovart's signature interaction: users modify design elements on the canvas in real time through natural-language instructions (changing copy, switching palettes, adjusting composition, adding/removing elements) rather than starting over. The infinite canvas accepts all outputs and supports layer separation and multi-round conversational editing.

The engineering significance of this design is that it converges the "generate → evaluate → correct" iteration loop onto a single state object (the canvas), avoiding cross-platform transfer and context loss — a direct embodiment of L3 orchestration's "state continuity".

3.3 Brand Kit (Brand Asset Kit)

Brand Kit is the platform's most distinctive design at the L4 layer. A Brand Kit usually contains:

The official tier-upgrade signal given in the brand-asset scenario is: first run a Pilot on the free tier, then upgrade only when SKU count rises — that is, tying "asset-reuse depth" to "subscription tier". The Brand Kit count cap is Starter 5 / Basic 10 / Pro 30 / Ultimate 100.

Third-party comparison figures (low confidence, reference only): traditional design companies quote roughly $3,000—10,000 for a brand identity kit, with a 2—4 week delivery cycle; Lovart produces a first draft within minutes on a subscription basis, but final output quality and deliverability still require human review.

3.4 Four Output Forms

3.5 Model Routing and Orchestration Objects

The platform explicitly takes "orchestrating multiple models" rather than "developing a single in-house model" as its technical route. The public model list and Hot Models unit prices at research time are as follows (from the official site's multilingual pages; prices change at any time, so the official real-time page must be taken as authoritative):

Publicly mentioned accessible models also include (with different figures across tiers): Nano Banana Pro, Seedream 4.0 / Seedance 1.5 Pro, Gemini Imagen 4, Recraft V3, Ideogram 3, Midjourney, Kling O1; on the video side, Sora, Veo, Kling, Wan, Vidu G2, Hailuo, LTXV 2.0, Runway Gen4, and more.

3.6 Team Collaboration and Delivery

• Each tier limits the number of concurrent tasks: Starter 2 / Basic 4 / Pro 8 / Ultimate 10.
• Pro and above target team and studio scenarios; third-party figures say they include team seats and collaboration capabilities ().
• Outputs can be exported to external editors such as Figma and Photoshop for further refinement.
• Every tier includes "100 refresh credits / day" and "access to all image/video/editing capabilities"; commercial licensing starts at the Basic tier.

4. Platform Architecture

图 4-1｜Lovart 平台总体架构（五层栈：交互 → 编排 → 资产 → 路由 → 计量）

数据来源：基于本文分析绘制的示意图。

4.1 Overall Architecture Layers

4.2 Inference and Orchestration Pipeline

The execution chain of a typical task is:

接收简报（自然语言）
  → 需求澄清（反问受众/调性/尺寸/语言）
  → 任务拆解（Logo / 主视觉 / 社媒组图 / 包装样机 / 分镜 …）
  → 为每个子任务选择模型（路由）
  → 批量并行生成（受并发配额限制：2 / 4 / 8 / 10）
  → 结果落到无边画布并完成排版
  → 用户以自然语言下达修改指令（Text Edit 或局部重生成）
  → 导出交付

A notable engineering detail: "modification" is split into two paths: Text Edit (targeted text changes) and Regenerate (regeneration). The former is higher-fidelity and lower-cost, a practical criterion for judging whether a design Agent is mature.

4.3 Metering and Quota System

The platform simultaneously uses three types of capacity constraints, differentiated by tier:

5. Harness Design

5.1 L1 Context Engineering Layer

Lovart's context is composed of three types of elements:

1. Brief: the natural-language requirement description, the primary context source.
2. Brand Kit: brand identity, color palette, and font system are injected long-term as structured constraint context, ensuring brand consistency across tasks.
3. Canvas state: existing outputs, layers, and historical modification instructions form session-level context, so multi-round revisions need not be re-described.

The advantage is that context is "asset-ized" — Brand Kit turns "brand consistency" from an adjective in a prompt into a constraint that the system can enforce. This follows the same idea as FLUX.2's precise hex brand-color control (see article 9), but Lovart elevates it to the asset-kit level.

Risk: context is assembled automatically by the Agent, and users lack visibility into and auditability of its content; when the Agent misunderstands the brief, the cost of correcting errors is higher than in a manual parameter-tuning mode.

Assessment: Strong (asset-ized context), but auditability is insufficient.

5.2 L2 Tools and Execution Layer

• The tool shape is "model + editing action": image generation, video generation, 3D generation, audio synthesis, HD upscaling, outpainting, localized repainting, Text Edit, background processing, export.
• Model routing is equivalent to registering external models as schedulable tools; this is isomorphic in spirit to Runway's MCP Server (see article 8), but Lovart's routing is closed to external developers — users cannot register new tools themselves.
• No public documentation on Function Calling, MCP, or open API for external systems was found ().

Assessment: Strong (complete tool coverage), but the toolset is closed and not extensible.

5.3 L3 Orchestration and Control Layer

This is the strongest layer of the platform within the six-layer model, and it is the fundamental reason the product exists.

Comparison and positioning against other platforms in this group:

• Meitu Design Studio Agent Teams (article 5): multi-agent division of labor and collaboration, targeting the full e-commerce visual chain, with the deepest engineering implementation and business metrics.
• ComfyUI (article 12): the orchestration output is a version-controllable JSON graph; users have full control but must build it themselves.
• Lovart: orchestration is managed by the Agent, zero learning curve for users, but the orchestration process is invisible, non-exportable, and not version-controllable.

Key shortcoming: orchestration artifacts (task decomposition plans, model-selection records, parameters) show no export or version-control capability. For teams that need reproducibility and auditability, this means "deliverables can be delivered, but the process cannot be traced" — in sharp contrast to ComfyUI, where "the workflow is a JSON graph that can be put under version control". In addition, recalling the "model combinations are dynamically adjusted with copyright and collaboration situations" mentioned in section 4, cross-time reproducibility is also questionable.

Assessment: Strongest (productized orchestration), but orchestration artifacts cannot be exported or version-controlled.

5.4 L4 Memory and State Layer

Brand Kit is the only first-class abstraction in this group of platforms that makes "brand assets" an explicit quota resource.

Comparison reference:

• Gemini / Nano Banana (article 4): the research report explicitly noted it has "no first-class abstraction for brand asset kits (compared with LOVART Brand Kit)" — precisely Lovart's relative advantage.
• Runway (article 8): at most 3 Brand Kits, far fewer than Lovart's Ultimate tier's 100.
• Kling (article 3): Element creation is person/object-level assets, a lower level than the brand asset kit.

Risk: the Xingliu service-shutdown incident (2026-10-10) exposed an asset portability risk at the L4 layer — the announcement explicitly states "each account only supports migration once", "newly added content after migration cannot be migrated again", "content on the original platform will be completely deleted and unrecoverable after service stops", and "Xingliu memberships and Lovart.art memberships are two independent systems that are not interoperable or convertible". This is the same class of structural risk as Miaoya Camera's "digital avatar assets locked within a single product, not portable" (article 6), except that Lovart at least provides a migration window.

Assessment: Strong (first-class Brand Kit), but a platform-level asset portability risk exists.

5.5 L5 Evaluation and Observation Layer

• Observation scope: credit consumption, concurrency usage, generation history.
• No official Eval Set, Golden Dataset, or regression set mechanism was found; design quality evaluation relies on users' subjective judgment and manual review.
• In external materials, the official side provides business outcome metrics (such as brand kit generation speed, cost comparison against design company quotes) rather than technical metrics of generation quality.
• Third-party reviews generally note that outputs are still "strong first drafts" requiring manual refinement in Figma / Photoshop / Recraft. This effectively indicates the platform itself has not established a quantifiable quality gate.

Assessment: Weak.

5.6 L6 Governance and Security Layer

• Commercial licensing is clearly tiered: the Free tier has no commercial license; Basic and above include full commercial copyright.
• No official implementation documentation was found for the platform's explicit labeling and implicit metadata labeling under the "Measures on Labeling AI-Generated Synthetic Content". [To be filled]
• No independent boundary policy was found for face-swapping / outfit-changing and use of real-person likenesses. [To be filled]
• Brand generation involves trademark-similarity risk: no public documentation was found for a conflict-check mechanism between AI-generated Logos and existing trademarks.
• Model routing relies on many third-party models; the chain of training-data copyright and output-rights is borne by each model provider, and no unified responsibility statement at the platform level was found.

Assessment: Medium-Weak (clearly tiered licensing, but labeling and content-safety mechanisms are not publicly documented).

5.7 Six-Layer Maturity Summary

6. Case Studies

6.1 Verifiable Public Data

6.2 Typical Use Cases

The following are typical workflows recorded by the official side and third parties (use-case descriptions, not business cases with performance data):

• Cold-start for a startup brand: a single brief produces Logo variants, color palette, font specifications, and initial social media materials for the website and fundraising materials.
• Freelancers / small studios: batch-produce recurring social media image sets, ad creative, and proposal visuals for multiple clients.
• Marketing teams: generate batches of posts, banners, and short-video assets from one brief, keeping visual consistency across assets.
• Concept stage for small agencies: use the Agent to accelerate concept exploration, while precise edits are still done in Adobe Firefly or vector editors.
• Product launch asset kit: social media image sets + packaging mockups + storyboards + short videos produced in one go.

6.3 Items Not Found

• Official brand or merchant customer cases with quantified effect data: none found.
• Platform's labeling implementation details under the "Measures on Labeling AI-Generated Synthetic Content": no official documentation found.
• Accurate membership pricing and tier benefits for Xingliu (domestic version): only third-party figures found, and the platform has stopped accepting new subscriptions.
• Whether the platform provides a public API or supports external tool registration (MCP / Function Calling): no official documentation found.
• Technical details of the Star-3 model (parameter count, training data, evaluation scores): none found.

7. Summary

7.1 Strengths

1. Strongest orchestration in this group: fully Agent-izes the design production chain from "brief → complete deliverable set", so users need not understand models or parameters.
2. Brand Kit is a first-class citizen: brand asset kits are explicitly modeled and quota'd by tier (5/10/30/100), the only platform in this research to do so.
3. Most complete output forms: image, video, 3D, and audio as one offering, covering the full design-delivery chain.
4. Converged canvas state: ChatCanvas + infinite canvas + layer separation means multi-round revisions need not restart, with iteration costs significantly lower than single-shot generative tools.
5. Clear cost-magnitude advantage: relative to traditional design companies' brand kit quotes ($3,000—10,000, 2—4 weeks), the subscription model has an order-of-magnitude advantage at the draft stage.
6. Clearly tiered commercial licensing: full commercial copyright from the Basic tier onward.

7.2 Disadvantages and Risks

1. Orchestration process non-exportable and not version-controllable: in contrast to ComfyUI's JSON workflows, the process cannot be traced or regression-verified.
2. Questionable reproducibility: model combinations are dynamically adjusted with copyright and collaboration situations, so the same instruction may produce different styles at different times.
3. Weak L5 evaluation layer: no Eval Set; outputs widely judged as "strong first drafts" that still need manual refinement.
4. The domestic Xingliu has ceased service: service stops on 2026-10-10, migration limited to once, membership systems non-interoperable and non-convertible — domestic teams must directly evaluate the Lovart.art Chinese version and fully recognize the platform-level asset portability risk.
5. Conflicting pricing figures: the same tier shows significant price differences across multilingual pages and third-party reviews (e.g., Pro tier $45 and $90 coexist); enterprise procurement must rely on the official real-time page.
6. Opaque governance mechanisms: labeling compliance, real-person likeness usage boundaries, and trademark-similarity checks are all undocumented publicly.
7. Niche controversy: third-party views hold that Lovart, after "killing the product manager", tends to drift toward a general Agent, and vertical depth may be diluted (low confidence, opinion only).

7.3 Applicability Boundaries

7.4 Selection Recommendations

• Startups and individual founders: recommended to run a Pilot on the free tier (about 100 refresh credits/day) to verify the Agent's understanding of your brand tone; upgrade to the Basic tier (with commercial license) once SKU count or asset needs rise.
• Teams needing brand consistency: prioritize Basic and above to get Brand Kit and commercial licensing; if you manage many brands/clients, choose a tier by Brand Kit quota (10 / 30 / 100) rather than credit volume.
• Enterprises with mature design processes: position Lovart as a "concept and first-draft accelerator"; final delivery still happens in your existing toolchain (Figma / Illustrator / Recraft); do not include it in formal production pipelines requiring audit and regression verification.
• Teams needing reproducible, version-controllable workflows: should turn to ComfyUI-family solutions (see article 12), or build it themselves with FLUX.2's fixed snapshot endpoint (see article 9).
• Domestic teams: Xingliu ceased service on 2026-10-10; please use the Lovart.art Chinese version directly and complete the work migration within the migration window (note: limited to once per account).

Information Gap Statement

1. Seriously conflicting pricing figures: Lovart's same tier shows significant price differences across multilingual official pages and third-party reviews (Starter $16 / $19; Basic $27 / $32; Pro $45 / $90; Ultimate $109 / $199), and credit-amount statements vary (e.g., Starter 2,000 vs. 1,500—2,000). Must rely on the official real-time pricing page; all data in this article is marked [To be verified].
2. Xingliu membership pricing: only the third-party figure "from about ¥59/mo" was found; the platform has stopped accepting new subscriptions, so it cannot be verified. [To be filled]
3. Version number: the platform has not published a unified version-naming scheme, so an accurate "latest version" cannot be given. [To be filled]
4. In-house Star-3 model: technical details (parameter count, training data, evaluation scores) not found.
5. API and external tool registration: no official documentation found on whether the platform provides a public API or supports MCP / Function Calling.
6. Team seats and collaboration: third-party figures say the Pro tier includes team seats and collaboration, not confirmed on official pages.
7. Compliance implementation under the "Labeling Measures": no official documentation found on explicit labeling and implicit metadata labeling. [To be filled]
8. Official customer cases and quantified effect data: none found, honestly noted in section 6.3. Figures like "conversion rate +57%" and "click-through rate +7.69%" were not adopted in this article because they were retrieved from other platforms and do not apply here.
9. Single-batch output of "about 40 assets": from third-party reviews, not official figures.

8. References

1. Lovart Official Website — Lovart (under LiblibAI). https://www.lovart.ai/
2. Lovart · Brand Kit feature page (includes per-tier credits, concurrency, Brand Kit counts, and Hot Models unit prices). https://www.lovart.ai/pt/features/brand-kit
3. Lovart · Editable Brand Kit feature page (includes monthly/annual subscription prices and benefit comparison). https://www.lovart.ai/it/features/create-editable-brand-kit-ai-agent
4. Xingliu official website (includes the 2026-10-10 shutdown and migration announcement). https://www.xingliu.art/
5. Xingliu AI Product Analysis — Qingyi Network (includes the Lovart / Xingliu relationship, Star-3 model, and four output forms). https://www.ra0.cn/?p=13081/
6. A single financing amount surpassing Manus — this AI company targets globalization — Securities Times, 2025-10-23 (includes Lovart DAU and annualized revenue figures). https://www.stcn.com/article/detail/3399818.html
7. Chen Mian — Baidu Baike (includes Lovart release date, ChatCanvas, and Beta test data). https://baike.baidu.com/item/%E9%99%88%E5%86%95/66304026
8. Chen Mian — Sogou Baike (includes the LiblibAI / Xingliu / Lovart product lines and funding history). https://baike.sogou.com/v10000726100.htm
9. Lovart Review 2026: AI Design Agent Pricing & Alternatives — AITrendTool (includes tier and credit figures; third-party review, Low-Medium confidence). https://aitrendtool.com/alternatives/lovart
10. AI Brand Kit Generator Pricing — Lovart (includes Brand Kit composition and tier comparison; third-party mirror site, Low-Medium confidence). https://lovart.pro/ai-brand-kit-generator-pricing
11. LiblibAI Introduction — Yingjipai (includes Xingliu AI membership pricing figures). https://www.yjpoo.com/site/997.html
12. "Measures on Labeling AI-Generated Synthetic Content" — issued by the CAC and three other departments on 2025-03-14, effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
13. Civil Code of the People's Republic of China, Article 1019 — National People's Congress, 2020. (Prohibits infringement of portrait rights by means such as forgery through information technology.)